Jamie Dimon wrote about AI for a page and used no numbers

Jamie Dimon’s letter to shareholders, published on 6 April 2026, is the most widely read document in banking that nobody is required to read. It contains figures on everything: branches opened, people hired, capital ratios, credit losses. The section on artificial intelligence contains none.

What the letter says

The claims are strong and entirely qualitative:

The importance of AI is real, and while I hesitate to use the word transformational, it is.

The pace of adoption will likely be far faster than prior technological transformations, like electricity or the internet.

AI will affect virtually every function, application and process in the company.

On employment he is more careful than the summaries suggest. AI “will definitely eliminate some jobs, while it enhances others”, and it “will create many jobs, some we can see today in cybersecurity and AI itself, and some we can’t see”. Then the sentence that carries actual operating content:

There is a possibility that AI deployment will move faster than workforce adaptation to new job creation.

And the posture: “We will not put our heads in the sand. We will deploy AI, as we deploy all technology, to do a better job for our customers (and employees).”

The omission is the story

JPMorgan is not a company waiting to see. At the company update in February 2026, CFO Jeremy Barnum budgeted about $19.8 billion of technology spend for the year, up 10 percent, and reported that the number of AI use cases in production had doubled. On the second-quarter earnings call in July, Dimon put the count at “almost 1,000 use cases today”, while adding that the ones that really matter are “probably 50”. His figure for the internal LLM Suite in February was 150,000 people using it every week.

None of that is in the letter. That is a choice, and Dimon has explained the logic for it elsewhere.

Asked in February what his scorecard for AI was, he pointed at the same LLM Suite, said its users think they are saving around four hours a week each, and then said the firm keeps that out of its own arithmetic: “we can measure it, but we don’t give it credit in terms of that, because it’s too vague.”

Which is the honest position, and it is worth stealing. Numbers about AI value are, at this stage, largely unfalsifiable, because an annual-value figure is a lot of estimated hours multiplied by a lot of assumed rates. Put one in a shareholder letter and you own it for a decade, whereas leaving it out keeps the ability to change your mind. A CEO with a legal department writes the version that survives.

Apply the same test to your own vendor conversations. A supplier who quotes a productivity percentage without naming the baseline, the measurement window and who did the counting is quoting a marketing number. It is not necessarily wrong. It is just not evidence, and it cannot be checked after the contract is signed.

The sentence to copy

Strip the rhetoric and one line has an actual implication: deployment may move faster than workforce adaptation.

That is a scheduling claim, not a philosophical one. It says the two things run on different clocks, the technology on a quarterly one and the people on an annual one, and that the gap between them is where the damage happens. Not damage in the science-fiction sense. Damage in the sense of a support team that shrank before anyone worked out what the remaining people are for, or a sales floor where the tooling changed twice while the compensation plan changed never.

The version of this that a 40-person company can act on has three parts, and none of them require a strategy document.

Name the affected work before you automate it. Not the job but the work. “Order confirmations” is work. “The person in the back office” is a job, and confusing the two produces the wrong conversation at the wrong time with the wrong person.

Put the adaptation on the same plan as the build. If a process goes live in six weeks, the question of what those hours become has a date too. Left undated, it resolves itself badly: the freed time gets absorbed invisibly, nobody can show a result, and the project is judged a failure by the same people who benefited from it.

Say what is not being cut. The single cheapest thing a managing director can do at the start of an automation project is state plainly whether it is a headcount project. Left unsaid, everyone assumes yes, and the assumption produces exactly the quiet non-cooperation that makes it a failure. We have watched this decide whether a pilot ever reaches productionDE (auf Deutsch) more often than any technical factor.

The letter’s AI section: zero figures. What Dimon said on the earnings call three months later: almost 1,000 AI use cases, and discrete areas where jobs were cut by 30% or 40%.

Letter to Shareholders, Annual Report 2025; 2Q26 Earnings Call Transcript, 14 July 2026

The numbers he gave later

The letter held that line for about three months. On the second-quarter earnings call on 14 July 2026, asked whether a bank could eventually run with a leaner structure, Dimon said the firm is “preparing to make sure we can retrain our people, and we have had discrete areas where we did reduce jobs by 30% or 40%”. Most of those people, he said, “were offered jobs elsewhere”.

That is the pattern rather than a contradiction. The figure arrived in a quarterly call, where it describes discrete areas after the fact and commits the firm to nothing forward, and not in the document a CEO is held to for a decade. The same answer carried the caveat that makes such a figure hard to bank: “you don’t uniquely benefit from AI. The ultimate beneficiary of AI will be our customers.”

Where the analogy breaks

Dimon compares the pace of adoption to electricity and the internet, and notes himself that those “took decades to roll out”. The comparison still flatters everyone who repeats it, because in both cases the lag was not technological. Factories had electric motors for years before anyone rearranged the factory around them; the productivity showed up when the layout changed, not when the motor arrived.

That is the more useful half of the analogy, and it argues against urgency in tool purchasing and for urgency in process design. Buying the model is the fast part. Rearranging the factory is the slow part, and it is the only part that produces a number worth putting in a letter.

A bank with the 300,000 employees Dimon counts can afford to run both clocks at once. A company with forty can afford one project at a time, which means picking the process where the layout actually changes, rather than the one where a tool is easiest to install.

The read

The letter is usually filed under “even Dimon says AI is transformational”. Read it for what it withholds instead: a man running one of the largest AI deployments in banking put no figure in it, while stating plainly that the technology may outrun the people. The figures came later, in a quarterly call, where they cost less to give.

Both halves of the letter are honest. Together they describe a company that is moving fast and refusing to pretend it knows the outcome, which is a more defensible position than most of what is said about AI at that altitude, and a better model for how to talk about it inside your own company than the confident version.

Frequently asked questions